visual-ai-assertions 0.20.0 → 0.22.0

package/dist/index.d.cts

@@ -32,6 +32,7 @@ declare const Provider: {
 /** Known model names grouped by provider. */
 declare const Model: {
     readonly Anthropic: {
+        readonly FABLE_5_1: "claude-fable-5-1";
         readonly FABLE_5: "claude-fable-5";
         readonly OPUS_5: "claude-opus-5";
         readonly OPUS_4_8: "claude-opus-4-8";
@@ -42,6 +43,7 @@ declare const Model: {
         readonly HAIKU_4_5: "claude-haiku-4-5";
     };
     readonly OpenAI: {
+        readonly GPT_6_ASTRA: "gpt-6-astra";
         readonly GPT_5_6_SOL: "gpt-5.6-sol";
         readonly GPT_5_6_TERRA: "gpt-5.6-terra";
         readonly GPT_5_6_LUNA: "gpt-5.6-luna";
@@ -84,7 +86,7 @@ type KnownModelName = (typeof Model.Anthropic)[keyof typeof Model.Anthropic] | (
 /** Default model selection used when a caller omits `config.model`. */
 declare const DEFAULT_MODELS: {
     readonly anthropic: "claude-sonnet-4-6";
-    readonly openai: "gpt-5.4-mini";
+    readonly openai: "gpt-5.6-luna";
     readonly google: "gemini-3-flash-preview";
     readonly openrouter: "qwen/qwen3.6-flash";
 };
@@ -667,6 +669,17 @@ interface VisualAIConfig {
      * Anthropic (Claude auto-downscales images).
      */
     imageDetail?: ImageDetailLevel;
+    /**
+     * Per-request timeout in milliseconds, forwarded to the provider SDK.
+     * Omitted by default, so each SDK's own default applies (OpenAI and
+     * OpenRouter 10 minutes, Google 1 minute, Anthropic per its own rules).
+     *
+     * Worth setting for heavy reasoning models: they can spend many minutes on a
+     * single call, and the SDK default lets a request hang far longer than most
+     * test suites should tolerate. Note that provider SDKs retry timed-out
+     * requests, so total wall time can exceed this value by a multiple.
+     */
+    timeout?: number;
     trackUsage?: boolean;
 }
 /** Optional instructions for `check()`. */
@@ -703,6 +716,37 @@ interface CompareOptions {
 /** Optional instructions for `elementsVisible()` and `elementsHidden()`. */
 interface ElementsVisibilityOptions {
     instructions?: readonly string[];
+    /**
+     * Whether the screenshot shows the interface in its finished state. Defaults
+     * to `true`, which is what a test asserting on a settled screen wants: an
+     * element under a loading spinner, skeleton or error overlay is reported as
+     * not properly visible, because the overlay says the content is not ready.
+     *
+     * Set to `false` when the screenshot was deliberately captured mid-load, so
+     * loading chrome is expected rather than a defect. The finished-state rules
+     * are then left out of the prompt entirely, and only presence and clipping
+     * are judged. Note this omits the guidance rather than inverting it — to have
+     * the model actively disregard loading indicators, say so in `instructions`.
+     */
+    finalState?: boolean;
+    /**
+     * Whether an element that is present but clearly badly rendered should fail.
+     * Defaults to `false`: `elementsVisible()` is a presence check, and an element
+     * counts as visible if it is there at all, whatever it looks like.
+     *
+     * Set to `true` to judge presentation as well. Text at contrast too low to
+     * read, elements overlapping, an element out of alignment with its siblings,
+     * or text cut off mid-word then fail, with the model told to say the element
+     * is present before naming the defect. Opt in per assertion where a rendering
+     * defect is the thing being checked: measured with the rule on and off, it
+     * caught exactly the bullet naming a present-but-overlapping element and
+     * nothing else, while making the better model noticeably flakier on plain
+     * presence questions.
+     *
+     * Ignored by `elementsHidden()`, where the question is absence and a rendering
+     * defect cannot change the answer.
+     */
+    requireCorrectRendering?: boolean;
 }
 /** Options for the built-in accessibility template. */
 interface AccessibilityOptions {

package/dist/index.d.ts

@@ -32,6 +32,7 @@ declare const Provider: {
 /** Known model names grouped by provider. */
 declare const Model: {
     readonly Anthropic: {
+        readonly FABLE_5_1: "claude-fable-5-1";
         readonly FABLE_5: "claude-fable-5";
         readonly OPUS_5: "claude-opus-5";
         readonly OPUS_4_8: "claude-opus-4-8";
@@ -42,6 +43,7 @@ declare const Model: {
         readonly HAIKU_4_5: "claude-haiku-4-5";
     };
     readonly OpenAI: {
+        readonly GPT_6_ASTRA: "gpt-6-astra";
         readonly GPT_5_6_SOL: "gpt-5.6-sol";
         readonly GPT_5_6_TERRA: "gpt-5.6-terra";
         readonly GPT_5_6_LUNA: "gpt-5.6-luna";
@@ -84,7 +86,7 @@ type KnownModelName = (typeof Model.Anthropic)[keyof typeof Model.Anthropic] | (
 /** Default model selection used when a caller omits `config.model`. */
 declare const DEFAULT_MODELS: {
     readonly anthropic: "claude-sonnet-4-6";
-    readonly openai: "gpt-5.4-mini";
+    readonly openai: "gpt-5.6-luna";
     readonly google: "gemini-3-flash-preview";
     readonly openrouter: "qwen/qwen3.6-flash";
 };
@@ -667,6 +669,17 @@ interface VisualAIConfig {
      * Anthropic (Claude auto-downscales images).
      */
     imageDetail?: ImageDetailLevel;
+    /**
+     * Per-request timeout in milliseconds, forwarded to the provider SDK.
+     * Omitted by default, so each SDK's own default applies (OpenAI and
+     * OpenRouter 10 minutes, Google 1 minute, Anthropic per its own rules).
+     *
+     * Worth setting for heavy reasoning models: they can spend many minutes on a
+     * single call, and the SDK default lets a request hang far longer than most
+     * test suites should tolerate. Note that provider SDKs retry timed-out
+     * requests, so total wall time can exceed this value by a multiple.
+     */
+    timeout?: number;
     trackUsage?: boolean;
 }
 /** Optional instructions for `check()`. */
@@ -703,6 +716,37 @@ interface CompareOptions {
 /** Optional instructions for `elementsVisible()` and `elementsHidden()`. */
 interface ElementsVisibilityOptions {
     instructions?: readonly string[];
+    /**
+     * Whether the screenshot shows the interface in its finished state. Defaults
+     * to `true`, which is what a test asserting on a settled screen wants: an
+     * element under a loading spinner, skeleton or error overlay is reported as
+     * not properly visible, because the overlay says the content is not ready.
+     *
+     * Set to `false` when the screenshot was deliberately captured mid-load, so
+     * loading chrome is expected rather than a defect. The finished-state rules
+     * are then left out of the prompt entirely, and only presence and clipping
+     * are judged. Note this omits the guidance rather than inverting it — to have
+     * the model actively disregard loading indicators, say so in `instructions`.
+     */
+    finalState?: boolean;
+    /**
+     * Whether an element that is present but clearly badly rendered should fail.
+     * Defaults to `false`: `elementsVisible()` is a presence check, and an element
+     * counts as visible if it is there at all, whatever it looks like.
+     *
+     * Set to `true` to judge presentation as well. Text at contrast too low to
+     * read, elements overlapping, an element out of alignment with its siblings,
+     * or text cut off mid-word then fail, with the model told to say the element
+     * is present before naming the defect. Opt in per assertion where a rendering
+     * defect is the thing being checked: measured with the rule on and off, it
+     * caught exactly the bullet naming a present-but-overlapping element and
+     * nothing else, while making the better model noticeably flakier on plain
+     * presence questions.
+     *
+     * Ignored by `elementsHidden()`, where the question is absence and a rendering
+     * defect cannot change the answer.
+     */
+    requireCorrectRendering?: boolean;
 }
 /** Options for the built-in accessibility template. */
 interface AccessibilityOptions {

package/dist/index.js

@@ -20,6 +20,7 @@ var Provider = {
 };
 var Model = {
   Anthropic: {
+    FABLE_5_1: "claude-fable-5-1",
     FABLE_5: "claude-fable-5",
     OPUS_5: "claude-opus-5",
     OPUS_4_8: "claude-opus-4-8",
@@ -30,6 +31,7 @@ var Model = {
     HAIKU_4_5: "claude-haiku-4-5"
   },
   OpenAI: {
+    GPT_6_ASTRA: "gpt-6-astra",
     GPT_5_6_SOL: "gpt-5.6-sol",
     GPT_5_6_TERRA: "gpt-5.6-terra",
     GPT_5_6_LUNA: "gpt-5.6-luna",
@@ -69,12 +71,16 @@ var Model = {
 };
 var DEFAULT_MODELS = {
   [Provider.ANTHROPIC]: Model.Anthropic.SONNET_4_6,
-  [Provider.OPENAI]: Model.OpenAI.GPT_5_4_MINI,
+  [Provider.OPENAI]: Model.OpenAI.GPT_5_6_LUNA,
   [Provider.GOOGLE]: Model.Google.GEMINI_3_FLASH_PREVIEW,
   [Provider.OPENROUTER]: Model.OpenRouter.QWEN_3_6_FLASH
 };
 var DEFAULT_MAX_TOKENS = 4096;
 var OPENAI_REASONING_MAX_TOKENS = 16384;
+var OPENAI_HEAVY_REASONING_MAX_TOKENS = 32768;
+var MODELS_REQUIRING_LARGE_OUTPUT_BUDGET = /* @__PURE__ */ new Set([
+  Model.OpenAI.GPT_6_ASTRA
+]);
 var MODEL_TO_PROVIDER = new Map([
   ...Object.values(Model.Anthropic).map((m) => [m, Provider.ANTHROPIC]),
   ...Object.values(Model.OpenAI).map((m) => [m, Provider.OPENAI]),
@@ -419,22 +425,50 @@ function buildComparePrompt(options) {
 }
 
 // src/templates/elements-visibility.ts
-var ELEMENTS_VISIBLE_ROLE = "Check whether specific UI elements are present and fully visible in this screenshot.";
 var ELEMENTS_HIDDEN_ROLE = "Check whether specific UI elements are absent or hidden in this screenshot.";
-var ELEMENTS_VISIBLE_EDGE_RULES = [
-  "If an element is partially visible (cut off by screenshot boundary), it is NOT considered fully visible \u2014 the check for that element should fail. Note the partial visibility in your reasoning."
+function joinClauses(clauses) {
+  if (clauses.length <= 1) return clauses[0] ?? "";
+  if (clauses.length === 2) return `${clauses[0]} and ${clauses[1]}`;
+  return `${clauses.slice(0, -1).join(", ")}, and ${clauses[clauses.length - 1]}`;
+}
+function visibleRole(finalState, requireCorrectRendering) {
+  const clauses = ["present", "properly visible"];
+  if (requireCorrectRendering) clauses.push("correctly rendered");
+  if (finalState) clauses.push("in their finished state");
+  return `Check whether specific UI elements are ${joinClauses(clauses)} in this screenshot.`;
+}
+var ELEMENTS_VISIBLE_CLIPPING_RULES = [
+  "When an element is partly rendered but cut off at an edge, decide whether ordinary scrolling would bring it fully into view. For example, a card peeking past the end of a horizontal carousel, a filter chip in a row that continues past the screen edge, or a list item partly below the bottom of a scrolling feed is reachable that way, so the check for that element PASSES. Say in your reasoning that it is reached by scrolling.",
+  "An element that scrolling cannot bring into view is NOT properly visible: one sliced by the screen edge itself, or cut off or overlapped by fixed chrome such as the status bar, a notch, a home indicator, a sticky header, or a fixed bottom navigation bar. That is a layout fault, so the check for that element FAILS. Describe the clipping in your reasoning.",
+  "An element you cannot see at all is not visible, even if the page might reveal it after scrolling. Judge only what this screenshot actually shows."
 ];
-var ELEMENTS_HIDDEN_EDGE_RULES = [
-  "If an element is partially visible (cut off by screenshot boundary), it is NOT considered hidden \u2014 the check for that element should fail. Note the partial visibility in your reasoning."
+var ELEMENTS_VISIBLE_FINAL_STATE_RULE = "Judge each element in its finished, presented state. Things a design draws on top of an element \u2014 a badge, a favourite icon, a duration or price pill, a gradient scrim \u2014 coexist with finished content and leave it visible. An overlay that says the element is NOT ready \u2014 a loading spinner, a skeleton placeholder, a shimmer, a progress bar, an error or retry overlay \u2014 means the element is not properly visible even when you can still make out what sits underneath, so the check for that element FAILS. Name which of the two you are seeing in your reasoning.";
+var ELEMENTS_VISIBLE_CORRECT_RENDERING_RULE = "An element that is present but clearly defective in how it is rendered is NOT properly visible: text at contrast too low to read, elements overlapping or colliding with one another, an element visibly out of alignment with the siblings it should line up with, or text cut off mid-word inside its own container. The check for that element FAILS. In your reasoning, say that the element is present and then name the defect. Only clear, unambiguous defects count: do not fail an element for tight spacing, stylistic choices, or anything you would have to argue for.";
+var ELEMENTS_VISIBLE_OCCLUSION_RULE = "An element a user could not read or use because a modal, dialog, cookie banner, toast or similar overlay covers it is NOT visible: the check for that element FAILS.";
+function visibleRules(finalState, requireCorrectRendering) {
+  return [
+    ...ELEMENTS_VISIBLE_CLIPPING_RULES,
+    ...finalState ? [ELEMENTS_VISIBLE_FINAL_STATE_RULE] : [],
+    ...requireCorrectRendering ? [ELEMENTS_VISIBLE_CORRECT_RENDERING_RULE] : [],
+    ELEMENTS_VISIBLE_OCCLUSION_RULE
+  ];
+}
+var ELEMENTS_HIDDEN_BASE_RULES = [
+  "An element that is rendered at all, even partly, is not hidden, so the check for that element FAILS. This includes one peeking past the edge of a scrollable row or feed, which the user reaches by scrolling normally. Note the partial visibility in your reasoning.",
+  "An element that appears nowhere in this screenshot counts as hidden, even if the page might reveal it after scrolling. Judge only what this screenshot actually shows."
 ];
+var ELEMENTS_HIDDEN_FINAL_STATE_RULE = "An element sitting under a loading spinner, skeleton, progress bar or error overlay is still rendered, so it is not hidden and the check for that element FAILS. It is not properly visible either; that is what the visible check is for.";
+function hiddenRules(finalState) {
+  return finalState ? [...ELEMENTS_HIDDEN_BASE_RULES, ELEMENTS_HIDDEN_FINAL_STATE_RULE] : ELEMENTS_HIDDEN_BASE_RULES;
+}
 function buildElementsVisibilityPrompt(elements, visible, options) {
-  const statements = visible ? elements.map((el) => `The element "${el}" is fully visible on the page`) : elements.map((el) => `The element "${el}" is NOT visible on the page`);
-  const defaultRules = visible ? ELEMENTS_VISIBLE_EDGE_RULES : ELEMENTS_HIDDEN_EDGE_RULES;
+  const statements = visible ? elements.map((el) => `The element "${el}" is visible on the page`) : elements.map((el) => `The element "${el}" is NOT visible on the page`);
+  const finalState = options?.finalState ?? true;
+  const correctRendering = visible && (options?.requireCorrectRendering ?? false);
+  const defaultRules = visible ? visibleRules(finalState, correctRendering) : hiddenRules(finalState);
   const instructions = options?.instructions ? [...defaultRules, ...options.instructions] : defaultRules;
-  return buildCheckPrompt(statements, {
-    role: visible ? ELEMENTS_VISIBLE_ROLE : ELEMENTS_HIDDEN_ROLE,
-    instructions
-  });
+  const role = visible ? visibleRole(finalState, correctRendering) : ELEMENTS_HIDDEN_ROLE;
+  return buildCheckPrompt(statements, { role, instructions });
 }
 
 // src/templates/accessibility.ts
@@ -544,6 +578,7 @@ function parseRetryAfter(value) {
 
 // src/providers/anthropic.ts
 var XHIGH_CAPABLE_MODELS = /* @__PURE__ */ new Set([
+  Model.Anthropic.FABLE_5_1,
   Model.Anthropic.FABLE_5,
   Model.Anthropic.OPUS_5,
   Model.Anthropic.OPUS_4_8,
@@ -567,12 +602,14 @@ var AnthropicDriver = class {
   maxTokens;
   apiKeyOrEnv;
   reasoningEffort;
+  timeout;
   constructor(config) {
     this.model = config.model;
     this.maxTokens = config.maxTokens;
     this.client = null;
     this.apiKeyOrEnv = config.apiKey;
     this.reasoningEffort = config.reasoningEffort;
+    this.timeout = config.timeout;
   }
   async getClient() {
     if (this.client) return this.client;
@@ -591,7 +628,10 @@ var AnthropicDriver = class {
         "Anthropic API key not found. Set ANTHROPIC_API_KEY or pass apiKey in config."
       );
     }
-    this.client = new Anthropic({ apiKey });
+    this.client = new Anthropic({
+      apiKey,
+      ...this.timeout !== void 0 && { timeout: this.timeout }
+    });
     return this.client;
   }
   async sendMessage(images, prompt, _options) {
@@ -689,6 +729,7 @@ var GoogleDriver = class {
   apiKeyOrEnv;
   reasoningEffort;
   imageDetail;
+  timeout;
   constructor(config) {
     this.model = config.model;
     this.maxTokens = config.maxTokens;
@@ -696,6 +737,7 @@ var GoogleDriver = class {
     this.apiKeyOrEnv = config.apiKey;
     this.reasoningEffort = config.reasoningEffort;
     this.imageDetail = config.imageDetail;
+    this.timeout = config.timeout;
   }
   toGeminiParts(images) {
     return images.map((img) => ({
@@ -719,7 +761,10 @@ var GoogleDriver = class {
         "Google API key not found. Set GOOGLE_API_KEY or pass apiKey in config."
       );
     }
-    this.client = new GoogleGenAI({ apiKey });
+    this.client = new GoogleGenAI({
+      apiKey,
+      ...this.timeout !== void 0 && { httpOptions: { timeout: this.timeout } }
+    });
     return this.client;
   }
   async sendMessage(images, prompt, _options) {
@@ -805,6 +850,7 @@ var OpenAIDriver = class {
   apiKeyOrEnv;
   reasoningEffort;
   imageDetail;
+  timeout;
   constructor(config) {
     this.model = config.model;
     this.maxTokens = config.maxTokens;
@@ -812,6 +858,7 @@ var OpenAIDriver = class {
     this.apiKeyOrEnv = config.apiKey;
     this.reasoningEffort = config.reasoningEffort;
     this.imageDetail = config.imageDetail;
+    this.timeout = config.timeout;
   }
   async getClient() {
     if (this.client) return this.client;
@@ -828,7 +875,10 @@ var OpenAIDriver = class {
         "OpenAI API key not found. Set OPENAI_API_KEY or pass apiKey in config."
       );
     }
-    this.client = new OpenAI({ apiKey });
+    this.client = new OpenAI({
+      apiKey,
+      ...this.timeout !== void 0 && { timeout: this.timeout }
+    });
     return this.client;
   }
   async sendMessage(images, prompt, options) {
@@ -903,6 +953,7 @@ var OpenRouterDriver = class {
   apiKeyOrEnv;
   reasoningEffort;
   imageDetail;
+  timeout;
   constructor(config) {
     this.model = config.model;
     this.maxTokens = config.maxTokens;
@@ -910,6 +961,7 @@ var OpenRouterDriver = class {
     this.apiKeyOrEnv = config.apiKey;
     this.reasoningEffort = config.reasoningEffort;
     this.imageDetail = config.imageDetail;
+    this.timeout = config.timeout;
   }
   async getClient() {
     if (this.client) return this.client;
@@ -928,7 +980,11 @@ var OpenRouterDriver = class {
         "OpenRouter API key not found. Set OPENROUTER_API_KEY or pass apiKey in config."
       );
     }
-    this.client = new OpenAI({ apiKey, baseURL: OPENROUTER_BASE_URL });
+    this.client = new OpenAI({
+      apiKey,
+      baseURL: OPENROUTER_BASE_URL,
+      ...this.timeout !== void 0 && { timeout: this.timeout }
+    });
     return this.client;
   }
   async sendMessage(images, prompt, options) {
@@ -1058,13 +1114,21 @@ function resolveConfig(config) {
 `
     );
   }
+  if (config.timeout !== void 0 && (!Number.isFinite(config.timeout) || config.timeout <= 0)) {
+    throw new VisualAIConfigError(
+      `Invalid timeout: ${config.timeout}. Must be a positive number of milliseconds.`
+    );
+  }
   const userSetMaxTokens = config.maxTokens !== void 0;
   let maxTokens = config.maxTokens ?? DEFAULT_MAX_TOKENS;
-  if (!userSetMaxTokens && (provider === "openai" || provider === "openrouter") && (config.reasoningEffort === "high" || config.reasoningEffort === "xhigh")) {
-    maxTokens = OPENAI_REASONING_MAX_TOKENS;
+  const effortNeedsLargeBudget = config.reasoningEffort === "high" || config.reasoningEffort === "xhigh";
+  const modelNeedsLargeBudget = MODELS_REQUIRING_LARGE_OUTPUT_BUDGET.has(model);
+  if (!userSetMaxTokens && (provider === "openai" || provider === "openrouter") && (effortNeedsLargeBudget || modelNeedsLargeBudget)) {
+    maxTokens = modelNeedsLargeBudget ? OPENAI_HEAVY_REASONING_MAX_TOKENS : OPENAI_REASONING_MAX_TOKENS;
     if (debug) {
+      const reason = modelNeedsLargeBudget ? `model "${model}", which exhausts smaller budgets on reasoning at any effort` : `provider "${provider}" with reasoningEffort "${config.reasoningEffort}"`;
       process.stderr.write(
-        `[visual-ai-assertions] Auto-increased maxTokens from ${DEFAULT_MAX_TOKENS} to ${OPENAI_REASONING_MAX_TOKENS} for provider "${provider}" with reasoningEffort "${config.reasoningEffort}".
+        `[visual-ai-assertions] Auto-increased maxTokens from ${DEFAULT_MAX_TOKENS} to ${maxTokens} for ${reason}.
 `
       );
     }
@@ -1077,6 +1141,7 @@ function resolveConfig(config) {
     reasoningEffort: config.reasoningEffort,
     maxImageDimension: config.maxImageDimension ?? DEFAULT_MAX_IMAGE_DIMENSION,
     imageDetail: config.imageDetail ?? DEFAULT_IMAGE_DETAIL,
+    timeout: config.timeout,
     debug,
     debugPrompt,
     debugResponse,
@@ -1087,6 +1152,11 @@ function resolveConfig(config) {
 // src/core/pricing.ts
 var PER_MILLION = 1e6;
 var PRICING_TABLE = {
+  // Fable 5.1 is priced identically to Fable 5, the tier it succeeds.
+  [`${Provider.ANTHROPIC}:${Model.Anthropic.FABLE_5_1}`]: {
+    inputPricePerToken: 10 / PER_MILLION,
+    outputPricePerToken: 50 / PER_MILLION
+  },
   [`${Provider.ANTHROPIC}:${Model.Anthropic.FABLE_5}`]: {
     inputPricePerToken: 10 / PER_MILLION,
     outputPricePerToken: 50 / PER_MILLION
@@ -1119,6 +1189,12 @@ var PRICING_TABLE = {
     inputPricePerToken: 1 / PER_MILLION,
     outputPricePerToken: 5 / PER_MILLION
   },
+  // Cached input is $1/MTok and cache writes $12.50/MTok; neither is modelled
+  // here, since `calculateCost` applies no cache discount on any provider.
+  [`${Provider.OPENAI}:${Model.OpenAI.GPT_6_ASTRA}`]: {
+    inputPricePerToken: 10 / PER_MILLION,
+    outputPricePerToken: 50 / PER_MILLION
+  },
   [`${Provider.OPENAI}:${Model.OpenAI.GPT_5_6_SOL}`]: {
     inputPricePerToken: 5 / PER_MILLION,
     outputPricePerToken: 30 / PER_MILLION
@@ -2127,10 +2203,34 @@ function stripCodeFences(text) {
 var CheckResponseSchema = CheckResultSchema.omit({ usage: true });
 var AskResponseSchema = AskResultSchema.omit({ usage: true });
 var CompareResponseSchema = CompareResultSchema.omit({ usage: true });
+var STRAY_CONTROL_CHARS = /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/g;
+function parseJson(text) {
+  try {
+    return JSON.parse(text);
+  } catch (first) {
+    try {
+      return JSON.parse(text.replace(/[\u0000-\u001f]/g, " "));
+    } catch {
+      throw first;
+    }
+  }
+}
+function stripControlCharacters(value) {
+  if (typeof value === "string") return value.replace(STRAY_CONTROL_CHARS, "");
+  if (Array.isArray(value)) return value.map((item) => stripControlCharacters(item));
+  if (value !== null && typeof value === "object") {
+    const out = {};
+    for (const [key, item] of Object.entries(value)) {
+      out[key] = stripControlCharacters(item);
+    }
+    return out;
+  }
+  return value;
+}
 function parseResponse(raw, schema) {
   let parsed;
   try {
-    parsed = JSON.parse(stripCodeFences(raw));
+    parsed = stripControlCharacters(parseJson(stripCodeFences(raw)));
   } catch {
     throw new VisualAIResponseParseError(
       `Failed to parse AI response as JSON: ${raw.slice(0, 200)}`,
@@ -2225,7 +2325,8 @@ function visualAI(config = {}) {
     model: resolvedConfig.model,
     maxTokens: resolvedConfig.maxTokens,
     reasoningEffort: resolvedConfig.reasoningEffort,
-    imageDetail: resolvedConfig.imageDetail
+    imageDetail: resolvedConfig.imageDetail,
+    timeout: resolvedConfig.timeout
   };
   const driver = createDriver(resolvedConfig.provider, driverConfig);
   const maxImageDimension = resolvedConfig.maxImageDimension;